Robust Face Recognition using Sparse Representation with a Novel Greedy Sparse Recovery Technique

This paper presents an efficient face recognition framework based on the Fast Reduced Sampling Matching Pursuit (FRSMP) algorithm. Unlike conventional greedy sparse recovery methods, which may select an inappropriate number of components per iteration, the proposed approach introduces an adaptive double threshold mechanism to determine relevant elements in a data-driven manner. Furthermore, it employs the Sherman Morrison Woodbury identity to iteratively update the least squares solution, thereby avoiding expensive large scale matrix inversions and significantly reducing computational complexity. The proposed method is extensively evaluated in terms of recognition accuracy, execution time, iteration count, and reconstruction error, and compared with state of the art techniques. When integrated into the kernel based sparse representation classification (KSRC) framework, FRSMP demonstrates notable performance gains, achieving high recognition rates (97.41% on LFW, 98.67% on GT, and 95.37% on AR) with reduced computational cost. The results confirm that the method outperforms existing sparse representation approaches while maintaining an effective balance between accuracy and elliciency, making it suitable for practical and resource constrained applications.